Ecology, 93(7), 2012, pp.

1593–1603
Ó 2012 by the Ecological Society of America

Effects of sampling completeness on the structure

of plant–pollinator networks
A. RIVERA-HUTINEL,1 R. O. BUSTAMANTE,2,3 V. H. MARÍN,2 AND R. MEDEL2,4
1
Instituto de Entomologı´a, Facultad de Ciencias Básicas, UMCE, Jose´ Pedro Alessandri 774, Santiago, Chile
2
Departamento de Ciencias Ecológicas, Facultad de Ciencias, Universidad de Chile, Las Palmeras 3425, Casilla 653, Santiago, Chile
3
Institute of Ecology and Biodiversity, Facultad de Ciencias, Universidad de Chile, Las Palmeras 3425, Casilla 653, Santiago, Chile

Abstract. Plant–animal interaction networks provide important information on commu-

nity organization. One of the most critical assumptions of network analysis is that the
observed interaction patterns constitute an adequate sample of the set of interactions present
in plant–animal communities. In spite of its importance, few studies have evaluated this
assumption, and in consequence, there is no consensus on the sensitivity of network metrics to
sampling methodological shortcomings. In this study we examined how variation in sampling
completeness influences the estimation of six network metrics frequently used in the literature
(connectance, nestedness, modularity, robustness to species loss, path length, and
centralization). We analyzed data of 186 flowering plants and 336 pollinator species in 10
networks from a forest-fragmented system in central Chile. Using species-based accumulation
curves, we estimated the deviation of network metrics in undersampled communities with
respect to exhaustively sampled communities and the effect of network size and sampling
evenness on network metrics. Our results indicate that: (1) most metrics were affected by
sampling completeness but differed in their sensitivity to sampling effort; (2) nestedness,
modularity, and robustness to species loss were less influenced by insufficient sampling than
connectance, path length, and centralization; (3) robustness was mildly influenced by sampling
evenness. These results caution studies that summarize information from databases with high,
or unknown, heterogeneity in sampling effort per species and should stimulate researchers to
report sampling intensity to standardize its effects in the search for broad patterns in plant–
pollinator networks.
Key words: accumulation curves; Clench model; ecological networks; Los Ruiles National Reserve,
Chile; network size; plant–pollinator network metrics; sampling completeness; sampling effort; sampling
evenness.

INTRODUCTION are often depicted in a matrix where columns and rows

The analysis of plant–animal interaction networks has
received considerable attention in recent years, in part
because it represents a tractable way to inquire into the
complex nature of community organization. By covering
such diverse topics as the relationship between biodi-
versity and stability (e.g., Memmott et al. 2004,
Bascompte et al. 2006), ecological generalization (e.g.,
Bascompte et al. 2003, Jordano et al. 2003), interaction
syndromes (e.g., Olesen et al. 2007), assembly rules (e.g.,
Guimarães et al. 2007, Santamaria and Rodrı́guez-
Gironés 2007), conservation (e.g., Fortuna and Bas-
compte 2006, Tylianakis et al. 2010, Gonzalez et al.
2011), and phylogenetic structure (e.g., Rezende et al.
2007), among others, the analysis of plant–animal
interaction networks has provided new insights in
ecology, evolution, and conservation of communities
(see review in Bascompte and Jordano 2007). Networks
Manuscript received 5 October 2011; revised 1 February
2012; accepted 17 February 2012. Corresponding Editor: R. E.
Irwin.
4 Corresponding Author. E-mail: rmedel@uchile.cl
1593
represent animal and plant species, respectively, and the
elements are the set of the realized interactions between
the two groups. Among the most commonly used
metrics to describe the structure of plant–animal
networks are connectance, nestedness, modularity, and
robustness (reviewed in Dormann et al. 2009). Exami-
nation of metrics, and therefore of network structure, in
different ecological systems and under different ecolog-
ical scenarios has produced a significant advance in the
understanding of general properties of plant–pollinator
networks. Nevertheless, an important shortcoming of
these metrics is their sensitivity to network size, that is,
to the total number of interacting species (animals plus
plants). This scale sensitivity implies that some network
properties do not necessarily remain constant with
changes in the number of species belonging to a network
(Bersier et al. 1999, Banasek-Richter et al. 2004). For
example, it has been described that, although modularity
and robustness are largely insensitive to network size
(e.g., Dunne et al. 2002, Olesen et al. 2007), an increase
in species richness tends to decrease connectance and to
increase nestedness (e.g., Jordano 1987, Bascompte et al.
1594 A. RIVERA-HUTINEL ET AL.
2003). In consequence, direct comparisons among
communities need to be interpreted with caution,
especially when network metrics are presented in
unstandardized form and network size effects are not
properly controlled (Jordano 1987, Bascompte et al.
2003, Olesen et al. 2006, 2007, Almeida-Neto et al.
2007). A second methodological shortcoming relates to
the sensitivity of network metrics to sampling intensity
(Blüthgen et al. 2006, Vázquez and Aizen 2006, Nielsen
and Bascompte 2007). Even though sampling effects are
not only circumscribed to network analyses, but also
pervade most empirical studies in ecology, conclusions
of the few studies that have addressed this issue suggest
that sampling intensity does not affect all network
metrics in the same way. For example, Nielsen and
Bascompte (2007) analyzed the sensitivity of connec-
tance and nestedness metrics to variation in sampling
effort, concluding that nestedness was less prone to
sampling bias compared to the number of species and
links in the network. Similarly, Petanidou et al. (2008)
showed that estimates of the degree of specialization of
animal and plant species depend strongly on sampling
effort, and tend to be inflated in studies with limited
sampling.
There are two broad approaches to examine the effect
of sampling bias on network structure. First, some
studies have focused on the analysis of fieldwork
sampling techniques and the extent to which conclusions
are influenced by the way samples are collected. For
example, Gibson et al. (2010) analyzed the potential bias
in network metric estimates when using time-based
observations or transects. Although connectance and
nestedness did not differ between the two sampling
procedures, time-based observations tend to record a
higher number of unique interactions and lower web
asymmetries than transect samplings. In the same vein,
Bosch et al. (2009) estimated network metrics from both
a phyto- and a zoocentric perspective, concluding that
analyses that rely on pollination visitation patterns to
record interactions often tend to underestimate con-
nectance and nestedness, compared to analyses that
identify interactions according to the pollen load
transported by pollinators. Second, data manipulation
after sampling permits one to evaluate whether network
statistics are susceptible to bias because of limited
sampling or confounding variables. In this regard, two
categories of analysis can be identified in the literature:
rarefaction analysis and accumulation curve analysis.
Rarefaction analysis consists of the random removal
of an increasing number of interactions in the network,
simulating a reduction in the number of interactions
recorded (e.g., Jordano 1987, Banasek-Richter et al.
2004, Blüthgen et al. 2006, Vázquez and Aizen 2006).
This is the only available method when researchers have
no direct access to the original data set. The procedure
has the disadvantage that it does not provide informa-
tion on observations that do not record any visit to a
plant, which we call failed sampling units (FSU,
Ecology, Vol. 93, No. 7
hereafter). Because by definition FSU are not incorpo-
rated into ecological network analysis (because only
interactions are recorded), the network contains neces-
sarily a fraction of the information collected. Because
FSU values provide valuable information on the time
spent by researchers to record interactions (Fig. 1), it is
necessary to include such events in estimates of potential
bias attributable to limited sampling effort.
Accumulation curve analysis consists of the recording
of new species (or interactions) for a community as a
function of increased sampling effort (e.g., Nielsen and
Bascompte 2007, Jordano et al. 2009, Chacoff et al.
2011). Unlike rarefaction analysis, this approach re-
quires direct access to the original data set. In the
context of network analysis, species accumulation curves
(or sample-based rarefaction curves; sensu Colwell et al.
2004) permit estimation of the expected number of
species or interactions as a function of the number of
observations (see Fig. 1c). When the accumulation curve
reaches an asymptote, it means that increasing sampling
effort will not increase the number of species or
interactions recorded (Colwell et al. 2004). To evaluate
sampling completeness, it is possible to calculate how far
the observed curve is from the expected asymptotic
value. Asymptote estimation may be performed by
fitting the observed accumulation curve to an asymp-
totic parametric model (Soberón and Llorente 1993), or
through a nonparametric estimator (e.g., Chao 2; see
Chacoff et al. 2011).
Accumulation curves may be useful descriptors of
sampling completeness for both species and interaction
levels. One shortcoming of accumulation curves based
on interactions, however, is that the number of
interactions can reach an asymptote even though some
plant species have been undersampled for pollinator
species. This phenomenon may occur when the sampling
effort is highly heterogeneous (i.e., low sampling
evenness). Poorly sampled plant species would thus be
associated with one or a few pollinator species,
overestimating the presence of specialists, inflating some
network metrics such as nestedness, and deflating others
such as connectance (Vázquez and Aizen 2006, Blüthgen
et al. 2008). This inconvenience may be overcome by
using accumulation curves on a per plant species basis to
evaluate the sampling completeness of the pollinator
assemblage visiting each plant species (e.g., Chacoff et
al. 2011). This analysis has some advantages: (1) it is
commensurate to the scale where observations are
performed (i.e., the plant species); (2) it permits
quantification of the heterogeneity in sampling effort
among plant species; and (3) it permits assessment of
levels of generalization–specialization among plant
species in relation to interspecific differences in visitation
rates (e.g., Herrera 2005). In this study we use species-
based accumulation curves to examine the sensitivity of
six widely used network metrics to variation in sampling
completeness per plant species in 10 plant–pollinator
networks. More specifically, we attempt to answer the
July 2012 SAMPLING EFFECTS ON MUTUALISTIC NETWORKS 1595

FIG. 1. Hypothetical example representing the overall procedure used in the study of plant–animal networks, and the
methodology used in this study. (a) Observed interactions between plant species (1, 2, 3) and animal species (A, B, C). The number
of time-based recordings for each interaction, including the null interactions (i.e., observations with no pollinator detected, 0’s), is
depicted in the first column. (b) Interaction matrix using the recordings in panel (a): the total numbers (R) of interactions for every
plant and animal species, and the total number of interactions recorded (m). Because null interactions (failed sampling; i.e., no visits
observed) are not included in the interaction matrix, the total sampling time (Ts) is not congruent with the number of interactions
recorded (m). (c) Species accumulation curve for plant species 1. The curve shows the number of pollinator species observed (So) at
increasing sampling effort. If the curve reaches an asymptote, it implies that the sampling effort is adequate to record the whole
pollinator assemblage. The Clench asymptotic model is shown by the dashed line (see Analyses of sampling effort for details). (d) By
adjusting the accumulation curve (So, lower curve) to the asymptotic models, it is possible to estimate the maximum number of
pollinators expected (Se, upper curve) and sampling completeness, that is, the fraction of the expected pollinator assemblage that is
recorded (So/Se).

following questions. (1) Does sampling completeness
influence network metric estimates? (2) Which metrics
are more robust to variation in sampling completeness?
(3) To what extent is the effect of sampling completeness
on network metrics influenced by the number of species
in the network and by sampling evenness?
MATERIALS AND METHODS
Fieldwork and sampling procedure
This study was conducted in an area near Los Ruiles
National Reserve (35850 0 S, 72830 0 W), a reserve aimed
to protect a rare and endemic deciduous forest in the
coastal range of central Chile. The Ruil forest is a highly
fragmented ecosystem distributed along 100 km of the
coastal range in central Chile (35–368 S latitude)
(Bustamante and Castor 1998). Currently, the landscape
is a mosaic in which the original forest has been almost
completely substituted by plantations of Pinus radiata
(Monterrey pine), an exotic tree originally from
California. The dominant species are the caducifolious
broad-leaved trees Nothofagus alessandrii and N. glauca
living with other subdominant evergreen trees such as
Cryptocarya alba, Aetoxicum punctatum, and Gevuina
avellana, among others (San Martin et al. 1984). Our
study focused on 10 fragments of native forest ranging
from 5.6 to 49.3 ha. Sampling of species and interactions
was performed during the austral spring and summer
seasons, when most flowering occurs in the study
system. Fieldwork was carried out during September
2007 through February 2008 and September 2008
through February 2009. On average, there were 85.8 6
11.8 plant species per fragment (mean 6 SD; range 67–
103) and 146.4 6 21.0 pollinator species per fragment
(mean 6 SD; range 113–189). Observations were
performed by 3.8 persons/day, on average, and always
on sunny days. The sampling procedure consisted of
recording the flower visitors to plants during a 10-min
1596 A. RIVERA-HUTINEL ET AL.
observation period (see Plate 1). Observations on plant
species were performed at random according to their
abundance in the study site. Overall, there were 1390 6
232 time-based observations per fragment (mean 6 SD;
range 1215–2004; n ¼ 10 observation periods), equiva-
lent to 231 hours of observation per fragment. Only
insect species that contacted the anthers or stigma of
flowers or entered the flower tube were considered as
legitimate visitors. We collected insects in the first year
of the study for subsequent taxonomic identification in
the laboratory. In total, we identified 186 flowering plant
species and 336 pollinator species in the overall system.
Analyses of accumulation curves and subsequent fitting
to the Clench asymptotic model revealed that our
sampling effort captured 96.9% of the expected plant
species richness (observed number of species, So ¼ 186;
expected number of species, Se ¼ 192; fit to Clench’s
model: r 2 ¼ 0.99) and 96.3% of the expected pollinator
species richness in the global system (So ¼ 336, Se ¼ 349;
r 2 ¼ 0.99).
Analyses of sampling effort
To investigate how sampling effort influences network
metrics, we used species-based accumulation curves for
the number of pollinator species visiting each plant
species. We calculated accumulation curves for every
plant species, including replicated analyses for the same
plant species in different sites. In total, we analyzed 687
species-based curves. For each accumulation curve we
estimated the mean and standard deviation from 100
random permutations of the data (Gotelli and Colwell
2001). The maximum number of pollinators at the
asymptote of each accumulation curve was estimated
through two different asymptotic models. (1) The
Clench model, a modification of the Michaelis-Mentel
equation, is defined as S(t) ¼ at/(1 þ bt); it is appropriate
when the probability of adding new species decreases
with the number of time-based observations (Soberón
and Llorente 1993). (2) The negative exponential model
is defined as S(t) ¼ (a/b)[1  ebt]. In the two models, S(t)
represents the predicted number of species at sampling
effort t (Fig. 1c), a is the rate of increase at the beginning
of the sampling, and b is a fitted constant that controls
the shape of the curve. In both cases, when t ! ‘, S(t)
! a/b. In consequence, the expected number of
pollinator species in the asymptote (Se) was calculated
as Se ¼ a/b (Soberón and Llorente 1993). To estimate the
parameters a and b, the functions must be adjusted to
the observed accumulation curves through a nonlinear
estimation procedure. We used the Gauss-Newton
algorithm for parameter estimations (Fletcher 2000).
The model that showed the best fit to the observed data,
as evaluated by Akaike’s information criterion, was
selected for subsequent analyses (Fox 2002).
Once we estimated the expected number of pollinators
per plant species at the asymptote, the sampling
completeness (SC) for each plant species was calculated
as SC ¼ So/Se, where So is the observed number of
Ecology, Vol. 93, No. 7
species and Se is the maximum number of species
expected according to the model (Fig. 1d). Based on
their sampling completeness, the plant species from each
site were assigned to one of five SC categories: 0–0.30,
0.31–0.50, 0.51–0.70, 0.71–0.90, and 0.91–1.00. For
example, subnetworks in the category 0.31–0.50 of
completeness include those plant species for which 31–
50% of the pollinator species expected by the asymptotic
model were recorded. The width of categories was
determined by finding the combination of ranges that
provided the lowest variation in the number of plant
species across categories. Finally, we calculated network
statistics for the set of species belonging to the same SC
category, totaling five subnetworks per fragment.
Because the same procedure was repeated for the 10
fragments here analyzed, our database consisted of 50
subnetworks, 10 per sampling completeness category. A
summary of network characteristics per sampling
category is shown in Table 1.
Colwell and Coddington (1994) reviewed diverse
estimators of species richness in samples, concluding
that parametric estimators are not necessarily indepen-
dent of the number of samples, especially in small
sample sizes. The authors found that Chao 2 and
Jackknife estimators provide the least biased estimators
for a small number of samples. To ensure that the
parametric procedure used in this study does not
underestimate species richness at small sample effort
categories, we examined the covariation of the expected
maximum species richness under Clench’s model and
Chao 2 procedure with ANCOVA, using sampling
completeness categories as treatment effect. All variables
were log-transformed. Results revealed an overall
association between estimates (F1, 384 ¼ 211.08, P ,
0.001, b ¼ 1.099 6 0.075, mean 6 SE) and lack of
sampling completeness effect (F4, 384 ¼ 1.558, P ¼ 0.185).
The slopes for the different sampling categories did not
differ (F4, 383 ¼ 0.927, P ¼ 0.448). Overall, these results
indicate that both procedures provide about the same
richness estimates, thus validating use of the parametric
procedure in this study.
Network metrics estimation and statistical analyses
We calculated the effect of sampling completeness on
six frequently used network metrics:
1) Connectance (C ), the proportion of possible links
actually observed in a web, with values from 0 to 1.
In bipartite networks it is calculated as C ¼ I/(P 3 A),
where I is the number of interactions, P is the number
of plant species, and A is the number of animal
species in the network. Connectance is often consid-
ered a measure of network complexity and redun-
dancy (Jordano 1987, Bascompte and Jordano 2006).
2) Nestedness (N ), representing one kind of asymmetry
of interactions, characterized by a core of highly
connected species (generalists) that interact mainly
with each other, and a group of specialist species that
July 2012 SAMPLING EFFECTS ON MUTUALISTIC NETWORKS 1597

TABLE 1. Network summary statistics (mean 6 2 SE) for the 50 subnetworks at Reserva Nacional Los Ruiles, Chile, 2007–2008.

Sampling completeness category

Network characteristics 0–0.30 0.31–0.50 0.51–0.70 0.71–0.90 0.91–1.00
Sampling completeness per species, SC 0.15 6 0.08 0.41 6 0.06 0.60 6 0.04 0.80 6 0.06 0.97 6 0.02
0
Sampling evenness, Jobs 0.88 6 0.03 0.95 6 0.04 0.93 6 0.03 0.92 6 0.03 0.88 6 0.03
Network size, S 78.0 6 13.3 36.9 6 6.3 49.6 6 8.5 81.0 6 13.8 83.8 6 14.3
Network asymmetry, NA 1.42 6 1.20 5.24 6 1.20 5.29 6 1.20 6.33 6 1.19 8.49 6 1.19
Number of plant species, P 32.4 6 1.2 6.9 6 1.5 8.0 6 1.2 10.8 6 1.3 10.0 6 1.9
Number of animal species, A 45.1 6 1.2 29.9 6 1.5 41.3 6 1.4 70.1 6 1.3 71.2 6 1.2
Number of interactions, I 63.2 6 4.5 92.2 6 8.3 93.7 6 6.6 105.7 6 8.1 93.7 6 7.2
Notes: N ¼ 10 subnetworks per sampling completeness category. The sampling completeness per species (SC), the fraction of the
expected pollinator assemblage that is recorded, was calculated from accumulation curves as SC ¼ So/Se, where So is the observed
number of species and Se is the expected number of species in the asymptote of Clench or exponential models. For example,
subnetworks in the category SC ¼ 0.31–0.50 include those plant species for which 31–50% of the pollinator species expected by the
0
asymptotic model were recorded. Sampling evenness (Jobs ) quantifies the heterogeneity of sampling effort among plant species.
0 0
When Jobs approaches 0, the effort distribution is highly heterogeneous among species; when Jobs approaches 1, the sampling effort
is homogeneously distributed among species. The network size (S ) within sampling completeness categories was calculated as the
sum of the plant (P) and animal (A) species belonging to the interval. Network asymmetry (NA) was calculated from the ratio (A/
P).

interact mainly with the generalist species (Bas-
compte et al. 2003). We used the NODF (nestedness
based on overlap and decreasing fills) algorithm
proposed by Almeida-Neto et al. (2008), which,
unlike previous metrics, corrects for matrix dimen-
sionality (Ulrich et al. 2009). NODF values of 0 and
100 indicate lack of nestedness and perfect nesting,
respectively.
3) Modularity (M ), providing information on the
existence of groups composed of strongly interactive
species that interact weakly with species belonging to
other modules (Guimerà and Amaral 2005). Diverse
algorithms have been suggested to estimate modu-
larity (see Olesen et al. 2007). Here, we used the
spinglass algorithm (Reichardt and Bornholdt 2006).
M can take values between 0 and ;1, and reaches its
maximum when all network interactions are within
modules and no link exists among modules.
4) Average path length (L), defined as the average
number of steps along the shortest paths for all
possible pairs of species in the network, or the fewest
number of steps required to connect species i with
species j, averaged for all pairs of species (Albert and
Barabasi 2002, Olesen et al. 2006). Although not
frequently used in studies of mutualistic networks
(but see Lundgren and Olesen 2005, Olesen et al.
2006, Bezerra et al. 2009), this metric provides
important information on network cohesiveness.
Low L values are associated with high levels of
species cohesion, a property characteristic of ‘‘small-
world’’ systems (Albert and Barabasi 2002).
5) Centralization (E), which permits identification of
central species around which peripheral species are
assembled. We calculated E as eigenvalue centrality
(Bonacich 1972), defined as the principal eigenvalue
of the adjacency matrix of a network, with values
between 0 and ‘. Eigenvector centrality assigns a
value to each species in the network. A species
receives a high E value if it is strongly correlated with
many other species that are themselves central within
the network. The presence of central nodes confers
high network cohesion or information flow (Borgatti
and Everett 1997).
6) Robustness (R), providing information on the
resilience of networks to losing species, based on
the concept of the attack tolerance curve (Albert and
Barabasi 2002). Curves are constructed by sequen-
tially removing plant or animal species without
replacement. After each species removal (e.g.,
pollinators), the number of species in the other
group (e.g., plants) that remain in the network is
calculated. The curve indicates the fraction of species
remaining according to the fraction of removed
species in the alternative group. The area under this
curve (values between 0 and 1) corresponds to the
resilience level of the overall system (Burgos et al.
2007). If the decay curve is concave (R , 0.5),
removing a small fraction of species produces a high
number of secondary extinctions. If the curve is
convex (R . 0.5), the network is highly robust. In
this study, we used a random species removal
criterion, and estimated independently the robust-
ness after plant removal (Rp) and animal removal
(Ra).
Because the sensitivity of metrics to sampling
completeness may vary depending on the number of
species in the network (network size) and sampling
evenness (the homogeneity of sampling effort among
plant species), we examined the pattern of covariation of
these variables with network metrics. The network size
within sampling completeness categories was calculated
as the sum of the plant and animal species belonging to
the interval. To quantify the heterogeneity of sampling
effort among plant species, we used an evenness index
0
(Jobs ) based on Shannon diversity (Blüthgen et al. 2008).
0
Jobs approaches 0 when sampling effort distribution is
highly heterogeneous, and approaches 1 when sampling
effort is homogeneously distributed among species. We
1598 A. RIVERA-HUTINEL ET AL.
TABLE 2. Summary of ANCOVA results for sampling completeness on network metrics.
Sampling completeness
Network metrics F4,43
Connectance, C 34.77***
Nestedness, NODF 26.67***
Modularity, M 30.87***
Robustness to plant loss, Rp 1.76
Robustness to animal loss, Ra 26.62***
Path length, L 26.62***
Centralization, E 43.53***
Notes: Network size and sampling evenness were used as covariates in every analysis. Symbols þ and  indicate positive and
negative patterns of covariation with sampling completeness, respectively. Symbol 0 indicates absence of significant covariation.
Network size, connectance, mean path length, and centralization values were log-transformed before analysis. Boldface indicates
values that retain significance after Bonferroni adjustment.
* P , 0.05; ** P , 0.01; *** P , 0.001.
analyzed the effect of sampling completeness on the
metric of interest with a one-way ANCOVA, using the
completeness category of 0.91–1.00 as reference for
comparisons and network size and sampling evenness as
covariates. A posteriori contrasts were performed using
Tukey HSD. All analyses were performed in R version
2.10 (R Development Core Team 2009) and associated
modules: car version 1.2–16 (Fox 2002), vegan version
1.15–4 (Oksanen et al. 2009), igraph version 0.5.3
(Csardi and Nepusz 2006) and bipartite version 1.06
(Dormann et al. 2008), all available online.5
RESULTS
The exponential asymptotic model provided the best
fit in 495 accumulation curves (72% of cases). The
number of observations per species was variable in the
overall system (mean: 26 time-based observations per
species, range: 1–283). In spite of this, when data were
analyzed at the subnetwork scale, the mean sampling
evenness, 0.91 6 0.05 (mean 6 SD), was greater than
those reported for other pollination networks (0.67 6
0.14, n ¼ 21 networks; see Blüthgen et al. 2008),
indicating that the species belonging to our sampling
completeness categories were more homogeneously
sampled than commonly reported. The mean complete-
ness of subnetworks did not relate to sampling evenness
(F1,48 ¼ 0.34, R 2 ¼ 0.007, P ¼ 0.560), suggesting that the
heterogeneity in the number of time-based observations
among species was associated with the distribution of
abundance rather than with variation in sampling effort
(Vázquez and Aizen 2006). Network size in the 50
subnetworks was 65.2 6 24.3 species (mean 6 SD), and
the animal/plant ratio (network asymmetry) was 5.35 6
2.92 (mean 6 SD). The A/P ratio (network asymmetry)
ranged from 1.4 in the category with the lowest sampling
completeness (0–0.30) to 8.5 in subnetworks containing
the best-sampled plant species (0.91–1.00), indicating
that well-characterized networks tend to present a
higher asymmetry in the number of interactive species
than low-sampled networks (see Table 1).
5 www.r-project.org
Ecology, Vol. 93, No. 7
Network size Sampling evenness
F1,43 Trend F1,43 Trend R2
13.57***  0.09 0 0.79
16.92*** þ 1.87 0 0.70
0.44 0 0.05 0 0.72
21.58*** þ 4.92* þ 0.63
0.86 0 5.24* þ 0.73
0.52 0 0.65 0 0.72
43.53*** þ 1.26 0 0.61
Connectance values differed among completeness
categories, showing an overall increase toward higher
values of sampling completeness (Table 2, Fig. 2a).
Subnetworks with species belonging to the three
categories with lowest completeness (ranging from 0 to
0.70) showed low connectance values and differed
statistically from the best-sampled subnetworks (Fig.
2a). Because by definition connectance values result
from the number of observed interactions relative to the
number of potential interactions (C ¼ I/P 3 A), it follows
that an increase in connectance will result from an
increased number of recorded interactions with respect
to the total number of species recorded. In fact, the
number of interactions (I ) increased with sampling
effort as expected (F1,48 ¼ 14.8, R 2 ¼ 0.22, P , 0.001).
Network size, i.e., the total number animal and plant
species within a network, was a significant covariate that
related inversely to network connectance (Table 2), as
shown in several analyses of plant–pollinator networks
(Bascompte and Jordano 2007). Sampling evenness, by
contrast, showed no relationship with connectance
(Table 2; P ¼ 0.766), indicating that this metric, one of
the most important descriptors of network structure,
was largely insensitive to the heterogeneity of sampling
effort among plant species.
Nestedness values differed among sampling categories
and tended to increase with sampling completeness (Fig.
2b). The only nestedness value differing from the
reference was observed in the SC category 0–0.3, where
nestedness was underestimated by 78% with respect to
the best-sampled category. Our findings are consistent
with previous conclusions that nestedness tends to
stabilize quickly with increasing sampling effort (Nielsen
and Bascompte 2007). In this study, nestedness values
increased with network size, as revealed by the
significant covariate effect (Table 2; P , 0.001).
Regarding sampling evenness, our results did not show
influence of heterogeneity of sampling effort on nested-
ness (Table 2). It has been suggested that incomplete and
heterogeneous samplings overestimate specialist interac-
tions and underestimate generalist interactions (Blüth-
gen et al. 2008, Dorado et al. 2011). However, our
July 2012 SAMPLING EFFECTS ON MUTUALISTIC NETWORKS 1599

FIG. 2. Network metric profiles across sampling completeness categories (the fraction of the expected pollinator assemblage
that is recorded, an indicator of sampling effort). All metric values are corrected by network size and sampling evenness. Bars
around means are 95% confidence intervals from 10 subnetworks contained in completeness intervals. Network metrics include (a)
connectance, (b) nestedness, (c) modularity, (d) network robustness to the loss of animal (solid circles) and plant (open circles)
species, (e) mean path length, and (f ) centralization. The values provided by the most intensively sampled category (0.91–1.00
completeness as revealed by species-based accumulation curves) were used as reference for statistical comparisons in one-way
ANCOVA.
* P , 0.05; *** P , 0.001.

results suggest that the NODF nestedness metric is
robust to heterogeneous sampling and shows deviations
only at very low sampling effort. In consequence, the
NODF metric seems to be more appropriate than
temperature-based nestedness, as suggested by Almei-
da-Neto et al. (2008) and Ulrich et al. (2009).
Mean modularity differed among categories and
showed a tendency to decrease with sampling complete-
ness (Table 2), which suggests that detection of
modularity in undersampled networks may represent
an artifact associated with insufficient sampling effort.
The only category showing a modularity level different
from the sampling reference was that in the lowest
sampling completeness (Fig. 2c). In principle, low-
sampled communities will record a limited number of
links, which may result in a pattern of low-connected
modules that inflates network modularity in comparison
to better-sampled communities. Neither network size
1600 A. RIVERA-HUTINEL ET AL. Ecology, Vol. 93, No. 7

PLATE 1. Plant–pollinator interaction between the solitary bee Manuelia postica (Apidae, Xilocopinae) and Olsynium junceum
(Iridaceae) at Los Ruiles Natural Reserve, Chile. Photo credit: R. Medel.

nor sampling evenness influenced modularity, as re-
vealed by the nonsignificant covariation of variables in
ANCOVA (Table 2; P ¼ 0.511 and P ¼ 0.824,
respectively). In general, modularity showed a robust
pattern to variation in sampling effort.
The robustness to animal loss (Ra ¼ 0.82 6 0.10, mean
6 SD) was greater than the robustness to plant loss (Rp
¼ 0.61 6 0.04) (t98 ¼ 13.69; P , 0.001; Fig. 2d),
indicating that animal and plant species removal had
different effects on network resilience. More specifically,
networks were more resilient when animals were
removed, in part because the number of animal species
exceeded the number of plant species in the overall
system (336 vs. 186 species, respectively), and the effect
of species loss may be more easily compensated in the
animal rather than plant species assemblage. Regarding
the effect of variation in sampling completeness, our
results indicate that, except for the low robustness to
animal loss detected at very low sampling effort (below
0.30 completeness), where networks were 23% less
resilient than the reference value (Fig. 2d), sampling
effort had a small impact on network resilience.
Network size influenced only the robustness to plant
loss (Table 2). The increased resilience of plant species to
increasing network size probably results from the
numerical asymmetry between animal and plant species.
If robustness estimates are influenced by the total
number of species used in analyses, networks composed
of a low number of plant species will show a low
robustness because only a few plant deletions are
required to remove many pollinator species, leading to
an overall system destabilization. In consequence, any
increase in the number of plant species will have a
disproportional effect on plant redundancy, turning the
network more robust to plant species loss. On the other
hand, if the high number of animal species implies a high
redundancy in the plant species with which they interact,
any increase in the number of animal species will have a
negligible impact on the already high network resilience.
Regarding sampling heterogeneity, our results did not
show a clear effect on network robustness. Although
both Rp and Ra tended to increase with sampling
evenness, the statistical significance disappeared after
Bonferroni adjustment (Table 2).
Mean path length (L) and centralization (E) metrics
were highly sensitive to sampling effort. Values of L and
E calculated from subnetworks containing species with
completeness lower than SC ¼ 0.70 differed significantly
July 2012 SAMPLING EFFECTS ON MUTUALISTIC NETWORKS
from the best-sampled subnetworks (Fig. 2e and f,
respectively). In general, increased sampling complete-
ness conferred high network cohesion, as revealed by the
decreasing mean path length and increasing centraliza-
tion values toward higher categories of sampling
completeness. Because more interactions are recorded
in well-sampled communities, species will be more
connected, on average, which implies that a lower
distance is necessary to connect pairs of interacting
species. Similarly, the greater number of interactions
recorded in well-sampled communities will increase the
importance of generalist species in determining network
cohesiveness. ANCOVA results indicate that centraliza-
tion, but not average path length, was influenced by
network size (Table 2), suggesting that the importance
of generalist species for network cohesion tends to
increase in large-sized communities. Neither path length
nor centralization was affected by sampling evenness.
DISCUSSION
Except for the robustness to plant loss, all of the
network metrics examined here were sensitive to some
extent to sampling effort. The most sensitive metrics
were connectance, path length, and centralization, which
gave underestimated values below sampling complete-
ness of 0.7 (Fig. 2a, e, f ). Nestedness, modularity, and
robustness to animal loss were the most robust metrics,
differing from well-sampled communities only in those
subnetworks with plant species showing ,30% of the
expected pollinator assemblage (Fig. 2b–d). Although
nestedness and robustness to animal loss showed
underestimated values in the lowest sampling interval,
modularity was overestimated by almost 40% (Fig. 2c).
Regarding network size effects, our results varied
depending on the network metric under assessment.
For example, connectance, robustness to animal loss,
and centralization showed high sensitivity to variation in
network size, but nestedness, modularity, robustness to
plant species loss, and path lengths were not affected.
Effects of sampling evenness were less prevalent in our
analysis. The only metrics influenced by the heteroge-
neity of sampling effort among plant species were those
related to robustness (Table 2). Specifically, our results
indicate that communities where sampling effort is more
homogeneously distributed among plant species will
show a higher robustness to species loss.
Our results assume that the most important cause
responsible for metric behavior across sampling effort
categories is sampling completeness per plant species.
However, subnetworks belonging to the same sampling
category may include species that share biological
properties such as phenology and floral characters,
among other variables, which might account, in part, for
the overall pattern. We examined this possibility by
comparing the similarity in species composition, as
estimated by Jaccard’s index, among and within
sampling categories. Subnetworks within sampling
categories were more similar in plant species composi-
1601
tion than those belonging to different sampling catego-
ries (ANOSIM test with 1000 permutations, F4,45 ¼ 2.24,
P , 0.001). To gain insight into the factors influencing
similarity in species composition, we used partial Mantel
tests to examine whether the correlation in the sampling
completeness distance was accounted for by sampling
effort (number of time-based observations per plant
species), taxonomic distance (phenotypic overlap), or
phenology distance (temporal overlap). The statistical
significance of such correlations was assessed by 1000
permutations. Our results indicate that variation in
sampling completeness was accounted for by sampling
effort (rm ¼ 0.327, P ¼ 0.001) but not taxonomy (rm ¼
0.012, P ¼ 0.759) and phenology (rm ¼ 0.001, P ¼
0.509). These results suggest that similarity in species
composition within categories was accounted for by the
relative abundance of species in the overall system,
which in turn influenced the sampling effort received by
each plant species, rather than by taxonomic and
phenological factors.
The aim of this study was to examine the extent to
which a number of frequently used network metrics are
influenced by sampling intensity, network size, and
sampling evenness. Although other studies have previ-
ously evaluated the effect of network size on the
structure of plant–animal relationships (e.g., Jordano
1987, Ollerton and Cranmer 2002, Blüthgen et al. 2006,
Vázquez and Aizen 2006, Almeida-Neto et al. 2007,
Nielsen and Bascompte 2007), studies examining the
contribution of different sources of bias to network
metric estimates have been notoriously absent in the
plant–pollinator literature. Even though species-based
accumulation curves have been used to examine the
impact of sampling effort on single network metrics
(e.g., nestedness as in Nielsen and Bascompte 2007) and
the level of generalization of individual plants in
populations of single species (e.g., Herrera 2005), this
is the first study using species-based accumulation curves
to evaluate the behavior of network descriptors to
variation in sampling effort in plant–pollinator net-
works.
Regarding the effect of network size on network
metrics, there is an important conceptual distinction
between our descriptive approach and scale–variance
analysis. It is one thing to state that a given metric is
statistically sensitive to variations in network size (or
any other aspect of network structure), which is often
examined using random matrices and investigating how
the values recorded for a given metric vary with network
size (e.g., Bersier et al. 1999, Vázquez and Aizen 2006,
Ulrich et al. 2009). It is another thing to state that a
given pattern has scale variance, which is often analyzed
by examining the statistical relationship of a metric with
network size at different ranges of network size, often
implying an underlying biological process (e.g., Marti-
nez 1992). In this work, we have described the statistical
sensitivity of network metrics to sampling effort,
keeping constant network size and sampling evenness.
1602 A. RIVERA-HUTINEL ET AL.
Because we do not know the statistical behavior of most
network metrics in response to network size, we cannot
ascribe causality to the network size dependency shown
by the metrics here analyzed (Table 2).
Results from this study have important consequences
for the way in which field studies on plant pollinator
should be designed. Because network metrics have
different sensitivities to sampling effort and network
size, studies having low sampling effort need to be
interpreted with caution. For instance, although nested-
ness, modularity, and robustness to species loss seem to
be robust estimators that show bias only at very low
sampling effort, conclusions on connectance, average
path length, and centralization may be more sensitive to
low sampling because they rely directly on the number
of interactions recorded. The lack of effect of sampling
evenness on most metrics suggests that heterogeneous
sampling effort across plant species is likely to have a
small influence on network metric biases. In conse-
quence, the reliability of network metrics may be
independent of whether sampling is homogeneous (i.e.,
a fixed an equal number of time-based observations per
plant; Lundgren and Olesen 2005) or random (i.e.,
recordings proportional to the relative abundance of
plant species, such as in this study) across species.
By identifying the network metrics most sensitive to
sampling effort, our results provide a baseline from
which to overcome the inherent sampling limitations of
most empirical studies. For instance, consideration of
additional variables such as the phenology of species
(e.g., Olesen et al. 2008, Petanidou et al. 2008), pollen
load (e.g., Bosch et al. 2009), pollen limitation (e.g.,
Power and Stout 2011), species abundance (e.g.,
Vázquez and Aizen 2006), invasion status of interacting
species (e.g., Aizen et al. 2008, Vilà et al. 2009), and
other factors may help to provide more complete
descriptions of plant–pollinator networks. We strongly
encourage researchers to document sampling effort
using accumulation curves at either network or species
levels. This information will permit standardization of
network metrics to variation in sampling procedures and
control of such effects in the search for broad patterns in
plant–pollinator networks.
ACKNOWLEDGMENTS
This study was funded by grants CONICYT-PBCT ACT 34/
2006 and PSD66. We thank J. San Martı́n, P. Peñailillo, and P.
Gomez for their help in plant identification. Insect identification
was performed with the invaluable help of D. Tapia, C.
Gonzalez, M. MacLean, M. Elgueta, and F. Rojas. CONAF
and Celco S.A. provided important facilities during fieldwork at
Los Ruiles National Reserve.
LITERATURE CITED
Aizen, M. A., C. L. Morales, and J. M. Morales. 2008. Invasive
mutualists erode native pollination webs. PLoS Biology
6:396–403.
Albert, R., and A. L. Barabasi. 2002. Statistical mechanics of
complex networks. Reviews of Modern Physics 74:47–97.
Almeida-Neto, M., P. Guimarães, P. Guimarães, Jr., R. Loyola,
and W. Ulrich. 2008. A consistent metric for nestedness
Ecology, Vol. 93, No. 7
analysis in ecological systems: reconciling concept and
measurement. Oikos 117:1227–1239.
Almeida-Neto, M., P. R. Guimarães, and T. M. Lewinsohn.
2007. On nestedness analyses: rethinking matrix temperature
and anti-nestedness. Oikos 116:716–722.
Banasek-Richter, C., M. F. Cattin, and L. F. Bersier. 2004.
Sampling effects and the robustness of quantitative and
qualitative food-web descriptors. Journal of Theoretical
Biology 226:23–32.
Bascompte, J., and P. Jordano. 2006. The structure of plant–
animal mutualistic networks. Pages 143–159 in M. Pascual
and J. Dunne, editors. Ecological networks: linking structure
to dynamics in food webs. Oxford University Press, Oxford,
UK.
Bascompte, J., and P. Jordano. 2007. Plant–animal mutualistic
networks: the architecture of biodiversity. Annual Review of
Ecology, Evolution, and Systematics 38:567–593.
Bascompte, J., P. Jordano, C. J. Melian, and J. M. Olesen.
2003. The nested assembly of plant–animal mutualistic
networks. Proceedings of the National Academy of Sciences
USA 100:9383–9387.
Bascompte, J., P. Jordano, and J. M. Olesen. 2006. Asymmetric
coevolutionary networks facilitate biodiversity maintenance.
Science 312:431–433.
Bersier, L., P. Dixon, and G. Sugihara. 1999. Scale-invariant or
scale-dependent behavior of the link density property in food
webs: a matter of sampling effort? American Naturalist
153:676–682.
Bezerra, E. L., I. C. Machado, and M. A. R. Mello. 2009.
Pollination networks of oil flowers: a tiny world within the
smallest of all worlds. Journal of Animal Ecology 78:1096–
1101.
Blüthgen, N., J. Fründ, D. P. Vázquez, and F. Menzel. 2008.
What do interaction network metrics tell us about special-
ization and biological traits? Ecology 89:3387–3399.
Blüthgen, N., F. Menzel, and N. Blüthgen. 2006. Measuring
specialization in species interaction networks. BMC [Bio Med
Central] Ecology 6:9.
Bonacich, P. 1972. Factoring and weighting approaches to
status scores and clique identification. Journal of Mathemat-
ical Sociology 2:113–120.
Borgatti, S. P., and M. G. Everett. 1997. Network analysis of 2-
mode data. Social Networks 19:243–270.
Bosch, J., A. M. Martı́n González, A. Rodrigo, and D.
Navarro. 2009. Plant–pollinator networks: adding the
pollinator’s perspective. Ecology Letters 12:409–419.
Burgos, E., H. Ceva, R. P. J. Perazzo, M. Devoto, D. Medan,
M. Zimmermann, and A. M. Delbue. 2007. Why nestedness
in mutualistic networks? Journal of Theoretical Biology
249:307–313.
Bustamante, R. O., and C. Castor. 1998. The decline of an
endangered temperate ecosystem: the ruil (Nothofagus
alessandrii ) forest in central Chile. Biodiversity and Conser-
vation 7:1607–1626.
Chacoff, N. P., D. P. Vázquez, S. B. Lomáscolo, E. L. Stevani,
J. Dorado, and B. Padrón. 2011. Evaluating sampling
completeness in a desert plant–pollinator network. Journal
of Animal Ecology 81:190–200.
Colwell, R., and J. Coddington. 1994. Estimating terrestrial
biodiversity through extrapolation. Philosophical Transac-
tions of the Royal Society of London B 345:101–118.
Colwell, R. K., C. X. Mao, and J. Chang. 2004. Interpolating,
extrapolating and comparing incidence-based species accu-
mulation curves. Ecology 85:2717–2727.
Csardi, G., and T. Nepusz. 2006. The igraph software package
for complex network research. InterJournal, Complex
Systems 1695. http://igraph.sf.net
Dorado, J., D. P. Vázquez, E. Stevani, and N. P. Chacoff. 2011.
Rareness and specialization in plant–pollinator networks.
Ecology 92:19–25.
July 2012 SAMPLING EFFECTS ON MUTUALISTIC NETWORKS
Dormann, C. F., B. Gruber, N. Blüthgen, and J. Fründ. 2009.
Indices, graphs and null models: analyzing bipartite ecolog-
ical networks. The Open Ecology Journal 2:7–24.
Dormann, C. F., B. Gruber, and J. Fründ. 2008. Introducing
the bipartite package: analysing ecological networks. R News
8:8–11.
Dunne, J. A., R. J. Williams, and N. D. Martinez. 2002.
Network structure and biodiversity loss in food webs:
robustness increases with connectance. Ecology Letters
5:558–567.
Fletcher, R. 2000. Practical methods of optimization. Second
edition. John Wiley, Padstow, UK.
Fortuna, M., and J. J. Bascompte. 2006. Habitat loss and the
structure of plant–animal mutualistic networks. Ecology
Letters 9:278–283.
Fox, J. 2002. An R and S-PLUS companion to applied
regression. Sage Publications, Thousand Oaks, California,
USA.
Gibson, R. H., B. Knott, T. Eberlein, and J. Memmott. 2010.
Sampling method influences the structure of plant–pollinator
networks. Oikos 120:822–831.
Gonzalez, A., B. Rayfield, and Z. Lindo. 2011. The disentan-
gled bank: how loss of habitat fragments and disassembles
ecological networks. American Journal of Botany 98:503–
516.
Gotelli, N. J., and R. K. Colwell. 2001. Quantifying biodiver-
sity: procedures and pitfalls in the measurement and
comparison of species richness. Ecology Letters 4:379–391.
Guimarães, P. R., G. Machado, M. A. M. de Aguiar, P.
Jordano, J. Bascompte, A. Pinheiro, and S. F. dos Reis. 2007.
Build-up mechanisms determining the topology of mutualis-
tic networks. Journal of Theoretical Biology 249:181–189.
Guimerà, R., and L. A. N. Amaral. 2005. Cartography of
complex networks: modules and universal roles. Journal of
Statistical Mechanics. Theory and Experiment, article
P02001.
Herrera, C. M. 2005. Plant generalization on pollinators:
species property or local phenomenon? American Journal of
Botany 92:13–20.
Jordano, P. 1987. Patterns of mutualistic interactions in
pollination and seed dispersal: connectance, dependence
asymmetries, and coevolution. American Naturalist
129:657–677.
Jordano, P., J. Bascompte, and J. M. Olesen. 2003. Invariant
properties in coevolutionary networks of plant–animal
interactions. Ecology Letters 6:69–81.
Jordano, P., D. Vázquez, and J. Bascompte. 2009. Redes
complejas de interacciones mutualistas planta–animal. Pages
17–41 in R. Medel, M. A. Aizen, and R. Zamora, editors.
Ecologı́a y evolución de interacciones planta–animal. Edito-
rial Universitaria, Santiago, Chile.
Lundgren, R., and J. M. Olesen. 2005. The dense and highly
connected world of Greenland’s plants and their pollinators.
Arctic, Antarctic, and Alpine Research 37:514–520.
Martinez, N. D. 1992. Constant connectance in community
food webs. American Naturalist 139:1208–1218.
Memmott, J., N. M. Waser, and M. V. Price. 2004. Tolerance
of pollination networks to species extinctions. Proceedings of
the Royal Society of London B 271:2605–2611.
1603
Nielsen, A., and J. Bascompte. 2007. Ecological networks,
nestedness and sampling effort. Journal of Ecology 95:1134–
1141.
Oksanen, J., R. Kindt, P. Legendre, B. O’Hara, G. L. Simpson,
P. Solymos, M. Henry, H. Stevens, and H. Wagner. 2009.
Vegan: community ecology package. R package version 1.15-
4. http://www.cran.r-project.org/package¼vegan
Olesen, J. M., J. Bascompte, Y. L. Dupont, and P. Jordano.
2006. The smallest of all worlds: pollination networks.
Journal of Theoretical Biology 240:270–276.
Olesen, J. M., J. Bascompte, Y. L. Dupont, and P. Jordano.
2007. The modularity of pollination networks. Proceedings of
the National Academy of Sciences USA 104:19891–19896.
Olesen, J. M., J. Bascompte, H. Elberling, and P. Jordano.
2008. Temporal dynamics in a pollination network. Ecology
89:1573–1582.
Ollerton, J., and L. Cranmer. 2002. Latitudinal trends in plant–
pollinator interactions: are tropical plants more specialised?
Oikos 98:340–350.
Petanidou, T., A. S. Kallimanis, J. Tzanopoulos, S. P.
Sgardelis, and J. D. Pantis. 2008. Long-term observation of
a pollination network: fluctuation in species and interactions,
relative invariance of network structure and implications for
estimates of specialization. Ecology Letters 11:564–575.
Power, E. F., and J. C. Stout. 2011. Organic dairy farming:
impacts on insect–flower interaction networks and pollina-
tion. Journal of Applied Ecology 48:561–569.
R Development Core Team. 2009. R: A language and
environment for statistical computing. R Foundation for
Statistical Computing, Vienna, Austria.
Reichardt, J., and S. Bornholdt. 2006. Statistical mechanics of
community detection. Physical Review E 74:016110.
Rezende, E. L., J. E. Lavabre, P. R. Guimarães, Jr., P. Jordano,
and J. Bascompte. 2007. Nonrandom coextinctions in
phylogenetically structured mutualistic networks. Nature
448:925–929.
San Martin, J., H. Figueroa, and C. Ramirez. 1984. Fitosocio-
logı́a de los bosques de ruil (Nothofagus alessandrii Espinosa)
en Chile. Revista Chilena de Historia Natural 57:171–200.
Santamaria, L., and M. A. Rodrı́guez-Gironés. 2007. Linkage
rules for plant–pollinator networks: trait complementarity or
exploitation barriers? PLoS Biology 5:354–362.
Soberón, J., and J. Llorente. 1993. The use of species
accumulation functions for the prediction of species richness.
Conservation Biology 7:480–488.
Tylianakis, J. M., E. Laliberte, A. Nielsen, and J. Bascompte.
2010. Conservation of species interaction networks. Biolog-
ical Conservation 143:2270–2279.
Ulrich, W., M. Almeida-Neto, and N. J. Gotelli. 2009. A
consumer’s guide to nestedness analysis. Oikos 118:3–17.
Vázquez, D. P., and M. A. Aizen. 2006. Community-wide
patterns of specialization in plant–pollinator interactions
revealed by null models. Pages 201–219 in N. M. Waser and
J. Ollerton, editors. Plant–pollinator interactions: from
specialization to generalization. University of Chicago Press,
Chicago, Illinois, USA.
Vilà, M., I. Bartomeus, A. C. Dietzsch, T. Petanidou, I. Steffan-
Dewenter, J. C. Stout, and T. Tscheulin. 2009. Invasive plant
integration into native plant–pollinator networks across
Europe. Proceedings of the Royal Society of London B
276:3887–3893.